Molecular Cell, Volume 60

Supplemental Information

A Regression-Based Analysis of Ribosome-Profiling Data Reveals a Conserved Complexity to

Mammalian Translation

Alexander P. Fields, Edwin H. Rodriguez, Marko Jovanovic, Noam Stern-Ginossar, Brian J. Haas,

Philipp Mertins, Raktima Raychowdhury, Nir Hacohen, Steven A. Carr, Nicholas T. Ingolia, Aviv Regev,

and Jonathan S. Weissman

Figure S1. Full distributions of ORF-RATER scores, related to Figures 1 and 2 and

Table 1.

(A) For each type of ORF, the cumulative distribution of the ORF-RATER score (i.e., the

number of ORFs receiving at least that score) is plotted. (B) The cumulative distribution

of the ORF-RATER score is plotted for each of the four codon species considered as

potential translation initiation sites. (C) The cumulative number of peptides captured by

ORF-RATER is plotted as a function of score threshold. If peptides could match multiple

CDSs, the one with the highest ORF-RATER score was selected. In addition to the

145,033 peptides documented in the plot, an additional 4,074 (for 149,107 total) were

identified that matched only CDSs not considered by ORF-RATER (e.g. due to low

sequencing coverage) but present in the UniProt mouse database. High-confidence

translated ORFs were defined as those receiving scores in the range 0.8–1, a conservative

threshold intended to limit the number of false positive identifications. Some bona fide

translated CDSs are not detected at this threshold, due to a low translation rate or

ambiguous ORF structure on their transcript(s) (e.g. multiple neighboring in-frame

AUGs). Additionally, some MS-detectable proteins may be missed by ribosome profiling

if they are no longer being translated.

Figure S2. Metagene profiles for novel CDSs following LTM or CHX treatment,

related to Figure 2.

Metagene profiles of each class of new CDS following treatment with LTM or CHX

display hallmarks of translation, such as peaks of density at translation initiation sites in

both datasets, and elevated average density in the body of CDSs following CHX

treatment.

Figure S3. A significant subset of CDSs expressed in both mouse BMDCs and HFFs

do not appear to be conserved for protein sequence, related to Figures 5 and 6.

(A) Cumulative distributions of per-codon PhyloCSF scores for the non-annotated

portions of all novel CDSs (left), uORFs (center), and extensions (right) identified as

translated in mouse BMDCs. CDSs for which no homologous translation initiation site

was identified in the HFF transcriptome received lower PhyloCSF scores; those for which

a homologous site was found in HFFs but not identified as translated received

intermediate PhyloCSF scores; and those for which the homologous site was identified as

translated in HFFs received greater PhyloCSF scores. Nonetheless, a significant fraction

of the CDSs translated in both mouse BMDCs and HFFs received negative PhyloCSF

scores, suggesting that the translation of those CDSs is conserved independent of the

encoded polypeptide sequence. (B) A uORF of Zdhhc3 encodes a peptide whose

sequence does not appear to be strongly conserved among Euarchontoglires, despite

being translated in both mouse BMDCs and HFFs. This uORF received a PhyloCSF

score of −101 decibans (−1.86 decibans/codon).

Figure S4. Fluorescent reporter constructs confirm translated CDSs of Zdhhc3 and

Fxr2, related to Figures 5 and 6. (A) Top, Zdhhc3 reporter construct design. The 5′ portion of the Zdhhc3 mRNA (ending

immediately prior to the first annotated transmembrane domain) was cloned from human

cDNA and fused to an in-frame eGFP sequence lacking its initial AUG. To enable

normalization for transfection and transcription efficiency, an mCherry sequence whose

translation was driven by an internal ribosome entry site (IRES) was included

downstream of the fused eGFP CDS. Point mutants of the Zdhhc3 gene sequence

corresponding to removal of all three of the uORFs, removal of the canonical translation

start site, and/or removal of the truncated translation start site were also generated (see

Methods for detail). Bottom, fluorescence measurements following transient transfection

of Zdhhc3 reporter constructs into HEK 293T cells. Removal of the uORFs increased

eGFP fluorescence, suggesting that translation of the uORFs inhibits downstream

translation. Removal of the canonical ATG increased eGFP fluorescence marginally,

whereas removal of the truncated translation start site abolished eGFP fluorescence,

suggesting that essentially all translation of the Zdhhc3 protein initiates at the

downstream ATG. Results for constructs lacking the uORF translation start sites in

addition to either the canonical or truncated ATGs suggest that even in the absence of

uORFs, the truncated ATG remains the preferred translation initiation site. (B) Top, Fxr2

reporter construct design. The reporters were identical to those constructed for Zdhhc3

(above) except that the sequence fused to eGFP was cloned from the human genomic

sequence for the first exon of Fxr2. The eGFP fluorescence level is most significantly

reduced following removal of the second of two upstream in-frame GTG codons

identified as a translation start site in mouse or human cells.

Figure S5. Peptides assigned to novel CDSs receive reasonable identification scores,

related to Figure 6. Tryptic peptides that match only newly identified CDSs score slightly more poorly than

peptides matching annotated proteins, but are much more reliably identified than those

matching reversed (decoy) peptides, both for MaxQuant (left) and Spectrum Mill (right).

Note that lower MaxQuant PEP scores (left) indicate a more-reliable identification,

whereas lower Spectrum Mill scores (right) indicate a less-reliable identification.

Figure S6. BC029722 encodes a conserved protein that localizes to the ER and Golgi

apparatus, related to Figure 6.

(A) A multiple sequence alignment of the protein encoded by the BC029722 gene

indicates that although the length of the protein varies, its N- and C-terminal regions are

well conserved among the Euarchontoglires. A homologous protein was not identified in

Rhesus macaque. (B) The human BC029722 homologue MMP24-AS1 fused to eGFP at

its C-terminus colocalizes with ER (top) and Golgi (bottom) markers when coexpressed

in HeLa cells. Scale bars, 10 μm.

Table S1. High-confidence translated ORFs identified by ORF-RATER in mouse

BMDCs, related to Figures 2 and 4. The 13,075 high-confidence translated CDSs ORF-RATER identifies in mouse BMDCs

are listed in this table. Genomic coordinates refer to the mm9 build of the mouse genome.

The “Homologous codon in HFF?” column indicates whether a corresponding NUG

codon could be identified in the HFF transcriptome; the “Translated codon in HFF?”

column indicates whether a high-confidence translation event was identified in HFFs that

initiated at that codon. “Novel peptides” indicates the number of peptides that could be

assigned to that CDS after excluding those that could have arisen from annotated

proteins. “PhyloCSF” is the direct output from the PhyloCSF algorithm (measured in

decibans) for only the portion(s) of the CDS non-overlapping with annotated CDSs, as

detailed in Methods. “PhyloCSF number of codons” indicates the size of the region to

which the PhyloCSF algorithm was applied. The final 26 columns are RPKM values for

each ORF at each time point before or during LPS stimulation. For these columns,

“CHX1” and “CHX2” refer to the two CHX-treated replicates, whereas “ND” was

collected in the absence of added drugs. The second part of the column name indicates

the number of hours of LPS treatment (e.g. 0h means untreated, 0.5h means half an hour

of treatment). Some ORFs could not be quantified due to their being too short. RPKM

values for very short or highly overlapping ORFs may be unreliable.

Table S2. Shotgun proteomics identifies peptides corresponding to some novel

translated CDSs, related to Figure 6

High-confidence seta

Extended setb

Proteins Peptides Proteins Peptides

Extensions 40 95 36 41

Isoforms 74 159 39 42

New 7 29 6 19

Truncations 38 46 76 87

Upstream 6 21 2 2

Total 165 350 159 191

a The high-confidence set includes ORFs assigned scores of 0.8–1 by ORF-RATER

b The extended set includes ORFs assigned scores of 0.5–0.8 by ORF-RATER

Table S3. Peptides identified from mouse DCs with their most parsimonious protein

source, related to Figure 6. The 149,107 tryptic peptides detected from mouse DCs are listed in this table. The first

column lists each peptide’s amino acid sequence. “MaxQuant score” and “MaxQuant

PEP” are the “Score” and “PEP” values reported by MaxQuant for those peptides it

identified. “Spectrum mill score” is the “score” reported by Spectrum Mill for those

peptides it identified. Not all peptides were identified by both programs. The following

six columns (“Gene”, “Chromosome”, “Start codon coordinate”, “Stop codon

coordinate”, “Strand”, “Protein length (AA)”, and “ORF type”) are as in Table S1. “Only

novel matches?” indicates whether the peptide could only have arisen from proteins

newly identified by ORF-RATER; a value of “FALSE” indicates that the peptide could

match an entry in UniProt, Ensembl, and/or UCSC KnownGene, even if this is not the

most parsimonious matching entry.

SUPPLEMENTAL EXPERIMENTAL PROCEDURES

BMDCs growth conditions

To obtain sufficient number of cells, we implemented a version of the BMDC

isolation protocol as described previously (Jovanovic et al., 2015). Briefly, 6–8 week old

female C57BL/6J mice were obtained from the Jackson Laboratories. RPMI medium

(Invitrogen) supplemented with 10% heat inactivated FBS (Invitrogen), β-

mercaptoethanol (50 μM, Invitrogen), L-glutamine (2 mM, VWR),

penicillin/streptomycin (100 U/mL, VWR), MEM non-essential amino acids (1X, VWR),

HEPES (10 mM, VWR), sodium pyruvate (1 mM, VWR), and GM-CSF (20 ng/mL;

Peprotech) was used throughout the study. At day 0, BMDCs were collected from femora

and tibiae and plated on twenty (per mouse), 100 mm non-tissue culture-treated plastic

dishes using 10 mL medium per plate. At day 2, cells were fed with another 10 mL

medium per dish. At day 5, cells were harvested from 15 mL of the supernatant by

spinning at 1,400 rpm for 5 minutes; pellets were resuspended with 5 mL medium and

added back to the original dish. Cells were fed with another 5 mL medium at day 7. At

day 8, all non-adherent and loosely bound cells were collected and harvested by

centrifugation. Cells were then resuspended with medium, plated at a concentration of

10×106 cells in 10 mL medium per 100 mm dish. At day 9, cells were stimulated for

various time points with LPS (100 ng/mL, rough, ultrapure E. coli K12 strain,

Invitrogen).

BMDC treatment for ribosome profiling

We generated two independent LPS time courses (9 time points: 0, 0.5, 1, 2, 4, 6,

8, 9, and 12 hours) and one Mock treated (LPS-free media) time course (same time

points) for BMDCs to be treated with CHX. We generated one LPS time course (same

time points) for ND treatment. The LTM-treated sample was an equal mix of BMDCs

stimulated with LPS at 8 of the 9 time points (all except 0.5-hour). Harr-treated samples

were unstimulated BMDCs (0-hour) and an equal mix of BMDCs stimulated with LPS

for 2, 4, 6, and 9 hours. We used 20 million cells for each single time point sample or

pooled sample.

For Harr inhibition of initiation, BMDCs were treated with 1 μg/mL Harr (LKT

Laboratories) at 37°C for 5 min. For LTM inhibition of initiation, BMDCs were treated

with 50 μg/mL LTM (Millipore) at 37°C for 30 min. For CHX inhibition of elongation,

BMDCs were treated with 100 μg/mL CHX (Sigma) at 37°C for 1 min. For all three

inhibitor treatments, cells were then treated with 100 μg/ml CHX at 37°C for 1 min, and

then washed twice with cold PBS containing 100 μg/mL CHX. For ND treatment, media

was removed from BMDCs, followed by flash freezing of the cells in liquid nitrogen. For

all of the treatments, cells were next lysed by triturating 10 times with 400 μL lysis buffer

(20 mM Tris pH 7.56, 150 mM NaCl, 5 mM MgCl2, 1% Triton X-100 [Sigma], 1 mM

DTT [Sigma], 8% glycerol, 100 μg/mL CHX, 12 units/mL Turbo DNase [Life

Technologies]). Lysates were clarified by centrifuging at 20,000×g for 10 min at 4°C.

Clarified lysate was flash frozen in liquid nitrogen.

BMDC protein isolation and processing for subsequent mass spectrometry

In order to capture proteins expressed at different times before or during LPS

stimulation, we collected 4 samples over the time course (0, 2, 6, and 12 hours). After

stimulation for the appropriate time, cells were washed twice with PBS and lysed for 30

min in ice-cold lysis urea buffer (8 M urea; 75 mM NaCl, 50 mM Tris HCl pH 8.0, 1 mM

EDTA, 2 μg/mL aprotinin [Sigma, A6103], 10 μg/mL leupeptin [Roche, #11017101001],

1 mM PMSF [Sigma, 78830]). Lysates were centrifuged at 20,000×g for 10 min, and

protein concentrations of the clarified lysates were measured via BCA assay (Pierce).

100 μg of total protein per time point were treated for 45 min with 5 mM dithiothreitol

(Thermo Scientific, 20291) to reduce protein disulfide bonds and alkylated for 45 min

with 10 mM iodoacetamide. Samples were then diluted 1:4 with 50 mM Tris HCl, pH

8.0, to reduce the urea concentration to < 2 M. Lysates were digested overnight at room

temperature with trypsin in a 1:50 enzyme-to-substrate ratio (Promega, V511X) on a

shaker. Tryptic peptides were desalted on C18 StageTips (Rappsilber et al., 2007) and

evaporated to dryness in a vacuum concentrator.

Desalted peptides were labeled with the iTRAQ reagent according to the

manufacturer’s instructions (AB Sciex) and as previously described (Mertins et al.,

2012). Briefly, 1 unit of iTRAQ reagent was used per sample (time point). 100 μg of

peptides were dissolved in 30 μL of 0.5 M TEAB pH 8.5 solution and the iTRAQ reagent

was added in 70 μL of ethanol. After 1 hour incubation the reaction was stopped with 50

mM Tris/HCl (pH 8.0). Differentially labeled peptides were mixed and subsequently

desalted on C18 StageTips (Rappsilber et al., 2007) and evaporated to dryness in a

vacuum concentrator.

To reduce peptide complexity and achieve deeper proteome coverage, samples

were separated by basic reversed-phase chromatography as previously described (Mertins

et al., 2013). Briefly, desalted peptides were reconstituted in 900 μL 20 mM ammonium

formate, pH 10, and centrifuged at 10,000×g to clarify the mixture before it was

transferred into autosampler tubes. Basic reversed-phase chromatography was conducted

on a Zorbax 300 Å Extend-C18 column, using an Agilent 1100 Series HPLC instrument.

The separations were performed on a 2.1 mm × 150 mm column (Agilent, 3.5 μm bead

size). Prior to each separation, columns were monitored for efficient separation with

standard mixtures containing 6 peptides. Solvent A (2% acetonitrile, 5 mM ammonium

formate, pH 10), and a nonlinear increasing concentration of solvent B (90% acetonitrile,

5 mM ammonium formate, pH 10) were used to separate peptides by their hydrophobicity

at a high pH. We used a flow rate of 0.2 mL/min and increased the percentage of solvent

B in a nonlinear gradient with 4 different slopes (0% for 1 min; 0% to 9% in 6 min; 9% to

13% in 8 min; 13% to 28.5% in 46.5 min; 28.5% to 34% in 5.5 min; 34% to 60% in 23

min; 60% for 26 min). Eluted peptides were collected in 96-well plates with 1 min (=

0.2 mL) fractions. Early eluting peptides were collected in fraction “A”, which is a

combined sample of all fractions collected before any major UV-214 signals were

detected. The peptide samples were combined into 24 to be used for proteome analysis.

Subfractions were achieved in a serpentine, concatenated pattern, combining eluted

fractions from the beginning, middle, and end of the run to generate subfractions of

similar complexities that contain hydrophilic as well as hydrophobic peptides. For high-

scale proteome analysis every 24th fraction was combined (1+25+49; 2+26+50; etc.).

Subfractions were acidified to a final concentration of 1% formic acid and desalted on

C18 StageTips (Rappsilber et al., 2007). LC-MS/MS analysis was performed as

previously described (Mertins et al., 2013). Each of the 24 subfractions was analyzed

twice on the LC-MS/MS in order to increase proteome coverage.

Ribosome profiling Ribosome-protected footprints were prepared for sequencing as previously

described (Stern-Ginossar et al., 2012). Briefly, clarified cell lysates were treated with

RNase I (Life Technologies) to digest RNA not protected by ribosomes. High molecular

weight species were isolated by centrifuging lysates through a 34% sucrose cushion at

200,000×g for 4 hours at 4°C. Pellets were resuspended in Trizol (Life Technologies) and

the RNA fraction isolated per the manufacturer’s instructions. LTM and ND samples

were treated with RiboZero (Epicentre Biotechnologies) after Trizol extraction, whereas

Harr and CHX samples were depleted of rRNA by subtractive hybridization. Total RNase

I-treated RNA was size-separated by gel electrophoresis, and RNAs of size 28–34

nucleotides were purified. These ribosome protected fragments were cloned and

sequenced via a single-end run on an Illumina HiSeq 2500 sequencer.

Transcriptome mapping of reads

Sequenced reads were first stripped of linker sequences used in the cloning

procedure. The resultant sequences were then filtered with Bowtie2 in --local mode for a

collection of rRNA, snoRNA, snRNA, miRNA, tRNA, mitochondrial rRNA, and

mitochondrial tRNA sequences compiled from UCSC and Ensembl annotations. The

remaining reads were then aligned with Tophat to a BMDC-specific transcriptome, which

was assembled from previously acquired RNA sequencing data (Shalek et al., 2013)

using the PASA platform (Haas et al., 2003). Reads not mapping to the BMDC-specific

transcriptome were then mapped with Tophat to the set of transcripts in UCSC Known

Gene (version of March 6, 2014) and Ensembl (version of March 11, 2014). The

following settings were used for Tophat alignments: --b2-very-sensitive --transcriptome-

only --no-novel-juncs --max-multihits=64.

P-site assignment Reads whose mapped positions overlapped annotated start codons were grouped

by drug treatment, read length, and presence of a mismatch at their 5′-most nucleotide

(reads with more than one such mismatch were excluded). For each of these groups, the

most frequent relative position of the adenine of the annotated start codon was assigned

as the P-site offset. The relevant P-site offset was then applied to all transcriptome-

mapped reads so that each read is mapped to a single transcript position.

Transcript selection Transcripts were removed from the collection if they included no mapped reads or

if the positions to which reads were assigned were entirely contained within a shorter

transcript. For each transcript, “multimapping” 29-nt subsequences identical to

subsequences on other transcripts at other genomic positions were identified, as well as

the number of 29- or 30-nt reads (in the ND dataset for the BMDCs, the CHX dataset for

the HFFs) aligning to each position. Transcripts were required to have at least 64 reads to

be kept, and transcripts for which 1/5 or more of the reads aligned to a single position

were also excluded. Transcripts overlapping annotated pseudogenes from

Ensembl/GENCODE (Pei et al., 2012) were excluded if more than 1/3 of their aligned

reads were in multimapping positions. Finally, the number of reads multimapping among

the remaining transcripts was calculated, and transcripts were removed if their fraction of

multimapped reads exceeded their fraction of multimapped positions by a value greater

than 1/3 (e.g., a transcript with 50% multimapping positions and 50% multimapping

reads would be kept, but one with 15% multimapping positions and 50% multimapping

reads would be removed).

ORF-RATER

Sets of transcripts sharing genomic positions on the same strand (i.e., RNA

isoforms) were grouped into “transcript families”. Reads aligning to the transcripts within

these families were assigned to the P-site as described. Metagene profiles were assembled

by averaging the read densities for one CDS per transcript family, selected as the one

found on the transcript with the greatest RNA-seq FPKM value. To be included in the

metagene, each CDS was required to include at least 50 aligned reads in all four drug-

treatment datasets. Prior to averaging, the read densities for each CDS were normalized

by their average value (across all of the treatments and footprint lengths). Metagene

profiles were prepared for each read length for each of the four treatments.

Transcript sequences were used to identify all NUG-initiated ORFs. For each

ORF, if at least one Harr or LTM read was within one nucleotide of the first base of the

initiator codon, profiles of expected read density were constructed from the metagene

read densities, explicitly modeling the positions from 3 nt upstream to 150 nt downstream

of the start codon (153 nt total) and from 18 nucleotides upstream to the end of the stop

codon (21 nt total). For ORFs shorter than 168 nucleotides, the explicitly modeled region

was truncated as necessary. For ORFs longer than 168 nucleotides, the remaining codons

were filled with 3 repeated values obtained by averaging the read densities of all codons

within the annotated CDSs, excluding the first 50 and last 7 codons, which were already

used to model the density near the start and stop codon. For each prospective start codon,

metagene profiles were also constructed to represent “abortive initiation” events,

consisting of only the read density on the initiator codon itself and the three nucleotides

upstream of it. For the CHX and ND datasets only, profiles were also constructed for

each stop codon, spanning the final codon and the stop codon, to account for the

possibility that a given stop codon might have especially slow kinetics

Following construction of the metagene profiles, the P-site mapped reads on each

transcript were fit using a non-negative least-squares regression (the nnls function of the

scipy.optimize Python library). Fits were performed first for the LTM and Harr datasets

independently; coefficient values from these regressions were summed for all ORFs

initiating at the same genomic coordinate (including the abortive initiation profile), and a

corresponding Wald statistic was calculated. Initiators for which no positive coefficients

were obtained were discarded from subsequent steps. Wald statistics were calculated

using a homoscedastic error model.

Next, non-negative least-squares regressions were performed on the combination

of the CHX and ND datasets. Coefficient values from these regressions were summed for

all ORFs initiating at the same genomic coordinate excluding the abortive initiation

profile, and a corresponding Wald statistic was calculated. Separately, coefficient values

were summed for all ORFs terminating at the same genomic coordinate, including the

stop-only profiles, and a corresponding Wald statistic was calculated. ORFs assigned a

regression coefficient value of zero were excluded from further analysis. Wald statistics

were calculated for ORFs aggregated by shared start or stop codon, rather than for each

individual ORF, because ORFs within those sets are frequently nearly (or actually)

linearly dependent, resulting in low confidence in the translation of any specific ORF but

high confidence that at least one of the set is translated.

Each genomic stop codon may serve as the terminator for multiple ORFs

initiating at different codons on the same or alternative transcript isoforms. To evaluate

the relative levels of translation initiating at each start codon within the group of ORFs

ending at the same stop codon, we assembled the summed regression coefficients for

each of those start codons and divided them by their maximum value.

For the purposes of evaluating translation, six features were collected for each

ORF: the three Wald statistics for its genomic start from the Harr, LTM, and CHX/ND

regressions; the Wald statistic for its genomic stop from the CHX/ND regression; the

relative translation level for its genomic start relative to the other ORFs terminating at the

same stop codon; and the actual magnitude of the summed regression coefficients for that

stop codon (to enable poorly translated ORFs to be penalized). Using these features, a

random forest consisting of 2048 trees was trained on all AUG-initiated ORFs at least

100 codons long, using the ORFs initiating and terminating at annotated start and stop

codons as the positive set and all of the other such ORFs as the negative set. Annotated

start and stop codons were taken from annotations in the Ensembl and UCSC Known

Gene transcript collections. The trees of the random forest were required to have at least

16 training examples in each leaf; this value was selected because it maximized cross-

validation accuracy among the set {8, 16, 32, 64, 128}. The random forest achieved 85%

six-fold cross validation accuracy on the training set and was next applied to all ORFs

regardless of length or start codon. We preferred a random forest over other machine

learning classifiers because random forests are simple and involve tuning only one free

parameter (the minimum number of training examples in each leaf; the algorithm is

relatively insensitive to the number of trees as long as a sufficiently large number is

used). Other classifiers, such as support vector machines, demand the imposition of a

distance metric, which we had no principled mechanism to select. A random forest, in

contrast, is sensitive only to each feature’s rank, not its magnitude.

All of the features used for classification by the random forest positively correlate

with likelihood of translation; however, the random forest does not guarantee that an

ORF with greater values for all features will necessarily receive a higher score. This can

lead to overfitting, in which sparsity of the training set results in some ORFs being

penalized despite having all higher feature values. To counteract this, we applied a

monotonization procedure, based on an equivalence to a network flow problem (Picard,

1976; Spouge et al., 2003), to assign a set of final scores that obeyed the monotonicity

constraint with minimum sum-of-squares deviation from the raw random forest scores.

Briefly, the feature values are used to construct a partial ordering of the ORFs, interpreted

as a directed acyclic graph, with each node corresponding to an ORF and directed links

indicating ordered pairs of ORFs for which the successor has feature values all greater

than or equal to those of the predecessor. The random forest score of each ORF is

associated as the “cost”. The transitive reduction of this graph (i.e., the minimal directed

graph preserving reachability) is then recursively partitioned: first, a “source” node is

connected to each of the nodes in the graph whose cost is above the average cost, with

edge weights set to the excess cost relative to the mean; a “sink” node is connected to the

other nodes, with each edge’s weight set to the difference between the mean cost and that

node’s cost; the weights of the internal edges are set to an effectively infinite value; the

minimum cut algorithm as implemented in the Python igraph library (Csardi and Nepusz,

2006) is applied to the graph, partitioning it into two subgraphs; and the process is

recursively applied to the subgraphs, terminating when one of the partitions contains no

ORFs but only the source or sink node. At this point, the score of each ORF within each

final partition is reset to the average score of all of the ORFs in that partition. High-

confidence translated ORFs were defined as those whose final score exceeded 0.8; the

extended confidence set was defined as those with score 0.5–0.8.

The ORF-RATER pipeline was implemented in Python and executed on a server

housing 64 cores (64-bit Intel Xeon X7560, 2.27 GHz) and 252 GB of RAM. Algorithms

were parallelized by chromosome. The slowest steps were the regressions against the

expected read profiles: the Harr regression took 7 hours on 11 cores; the LTM regression

took 4 hours on 6 cores; and the CHX/nodrug regression took 13 hours on 8 cores.

Including read alignment and all other steps, a typical dataset could be processed over the

course of a few days.

Novel CDS metagenes and phasing

Metagenes of novel translation events were compiled in much the same way as

for the annotated set. Only those novel ORFs long enough to encompass all of the codons

in the metagene plot were included in the metagene. The phasing values were obtained by

averaging together the reads in the ND dataset of length 29 and 30 in the appropriate

positions of each novel CDS (excluding start and stop codons and the codon preceding

them); these read lengths showed the strongest periodicity at canonical CDSs.

Quantification of translation

To quantify expression of the high-confidence CDSs within each transcript family

at each CHX or ND timepoint, simplified profiles were constructed for each in which the

same three values—the average fraction of reads in each coding frame across all of the

codons of canonical CDSs in the dataset in question—were assigned to each codon. For

each CDS, windows of three codons at the start and stop codons were excluded from the

analysis (the start codon and two surrounding codons, and the stop codon and two

preceding codons); reads of all lengths at the remaining codons were aggregated and a

regression was performed. In this way, the estimated translation at isolated CDSs is a

weighted sum in which reads in the proper reading frame contribute more than those in

other frames, and the estimated translation of overlapping CDSs is fractionally assigned

based on the read density in non-overlapping regions and based on the frame of the reads

in the overlapping region (if the CDSs are in different frames). The number of reads

assigned to any given CDS were then normalized by CDS length and sequencing depth to

produce reads per kilobase per million (RPKM) expression values for each CDS.

Translational dynamics and clustering

In order to investigate expression changes throughout the LPS stimulation time

series, we first median-centered the RPKM expression values of all CDSs for each

condition and time point. Because CHX and ND expression levels are well correlated, we

averaged the two CHX and single ND replicate for every CDS. These averaged values

were used for subsequent analysis. We required that all analyzed CDSs have at least a

sum of 10 averaged RPKM across the time series. For each CDS, a simple yet robust

translational “fold change” was defined as its maximum translation level divided by its

minimum translation level, after smoothing the translation time series using a sliding

window of three time points. For example, the fold change assigned to a CDS whose

translation monotonically declined was the average of the first three time points divided

by the average of the final three time points. The averaging procedure served to buffer

against outlying measurements. CDSs whose fold change exceeded a value of 2 were

selected for hierarchical clustering. Z scores were calculated for each CDS, and Pearson

correlation was used to quantify pairwise similarity (using the linkage function of the

scipy.cluster.hierarchy library). Based on these pairwise distances, 128 clusters of CDSs

were identified using the fcluster function of the scipy.cluster.hierarchy library.

Gene ontology analysis

From the 128 clusters formed by hierarchical clustering, 3 prominent examples

were chosen because they exhibit robust and distinct expression patterns, peaking in

expression early, mid, and late in the time series. For each of these three clusters, unique

annotated CDSs were identified and searched for enriched Gene Ontology Terms using

The Database for Annotation, Visualization, and Integrated Discovery (DAVID) (Huang

et al., 2008). We employed all Biological Process Terms (GOTERM_BP_ALL). To limit

redundancy in the final output, we made use of DAVID’s functional annotation

clustering, which groups GO terms based on similar gene members. We plot the top five

GO terms for each of the clusters identified in Figure 4D and their DAVID enrichment

scores, which reflect the geometric mean of the p values for all GO terms included in that

annotation cluster.

PhyloCSF

For each novel CDS, those codons for which no nucleotides overlapped annotated

ORFs were isolated. If at least ten such codons existed, corresponding sequences were

identified in a set of ten mammals spanning the Euarchontoglires and whose genome

assembly was in at least its second iteration: human, chimpanzee, rhesus macaque,

bushbaby, mouse, rat, guinea pig, squirrel, rabbit, and pika. Aligned sequences were

retrieved using the 60-way multispecies alignment available from the UCSC genome

browser. PhyloCSF software (Lin et al., 2011) was downloaded and applied to the

aligned sequences if they could be identified in at least five of the species. The

background distribution of PhyloCSF scores for uORFs were obtained by taking the set

of all AUG-initiated uORFs of high-confidence translated canonical CDSs (to avoid

uORFs on spurious transcripts) that were at least 10 codons long and non-intersecting

with any annotated CDSs. Similarly, the background distribution of PhyloCSF scores for

extensions were obtained for the set of all NUG-initiated extensions of translated CDSs

that spanned at least 10 codons, taking the minimal such extension for each CDS and

requiring no overlaps with any other canonical CDSs. The distribution of PhyloCSF

scores at intergenic regions were taken from a random sampling of AUG-initiated

intergenic ORFs of at least 10 codons.

Correspondence of BMDC and HFF CDSs

Following application of the ORF-RATER algorithm to both mouse and human

ribosome profiling datasets, liftover software (Hinrichs et al., 2006) was used to map the

genomic coordinates of the mouse translated initiator codons to the human genome. The

corresponding human initiator codon was defined as the nearest NUG up to a maximal

distance of 9 nucleotides; of the 10,634 translated mouse initiators, 9827 (92%)

corresponding human genomic positions were identified, of which 9056 (92%; 85%

overall) could be assigned to an initiator codon on a HFF transcript. Of those, 6186 (68%;

58% overall) had been identified by ORF-RATER as high-confidence translation

initiation sites in the HFF dataset. For translation initiation sites potentially leading to the

translation of multiple CDSs (on different transcript isoforms), the CDS length plotted in

Figure 5A was selected as the one assigned the highest ORF-RATER score. For the gene-

level analysis (Figure 5B), translation initiation sites for which a corresponding site could

be identified in HFFs were grouped for each family of mouse transcripts, and the total

number and the number for which a corresponding translation event in HFFs was

identified was calculated.

Identification of peptides and proteins

First, we generated a concatenated database including all canonical proteins in the mouse

UniProt database (July 2014), additional annotated proteins from Ensembl and UCSC

Known Gene that were missing from the UniProt database, and new proteins identified

by ORF-RATER (including scores 0.5 and above). The final search database included

64,997 entries. All mass spectra were analyzed with MaxQuant software version 1.5.2.8

(Cox et al., 2011; Cox and Mann, 2008) as previously described (Jovanovic et al., 2015),

and with Spectrum Mill software package v4.0 beta (Agilent Technologies, Santa Clara,

CA) as previously described (Mertins et al., 2012; Mertins et al., 2014). For both

software packages, we used default parameters and applied a maximum FDR of 1%

separately on the protein and peptide level. Allowed variable modifications followed

program defaults: methionine oxidation and N-terminal acetylation for both programs,

and asparagine deamidation to aspartate for Spectrum Mill. Only peptides uniquely

matching to newly identified proteins were considered to be positive identifications of a

novel translation event and included in Table S2. For Table S3, each peptide was

assigned to the most parsimonious protein from which it could have arisen. To make

these assignments, all potential protein assignments were identified for each peptide;

next, proteins were selected one by one that could explain the maximum number of

unassigned peptides, until all of the peptides had been assigned a source. If more than one

protein was available that could explain the same number of peptides, the one with the

maximal ORF-RATER score was selected. The MS dataset generated should be

considered of qualitative nature. Although it is a mix of BMDCs stimulated for different

length of time with LPS, the major goal was to increase proteome coverage, i.e., increase

the chance of detecting proteins during any phase of LPS stimulation, and not to provide

quantitative differences of protein expression between the time points.

Length distributions

Empirical nucleotide abundances were calculated from the set of mouse BMDC

transcripts. Based on these abundances, the frequency of stop codons was calculated to be

1/20. The distribution of ORF lengths on scrambled transcripts (Figure 3A) was

calculated as a geometric distribution with that parameter. The length distribution of

previously annotated CDSs was plotted based on the CCDS collection (Farrell et al.,

2014).

Protein multiple sequence alignments

The codon sequences of the proteins plotted in Figures S3B and S6A were

obtained for each mammal using the 60-way multispecies alignment, followed by

translation to amino acid sequence and alignment using Clustal omega (Sievers et al.,

2011). Alignments were plotted using Jalview (Waterhouse et al., 2009).

Cloning and cell culture

To generate the plasmid constructs used in Figure S4A, the first exon of the

human Fxr2 gene was amplified from purified human genomic DNA (courtesy M.

Leonetti) by PCR using primers 5′-CGA TTG ACT GAG TCG CCC GGA TCC GCA

GTA GGC GGC GGT G-3′ (forward) and 5′-CTC CTC GCC CTT GCT CAC ACC

AGA ACC ACC CTT GTA GAA GGC CCC GTT GG-3′ (reverse), where the underlined

segment is the portion complementary to the genomic DNA sequence. Similarly, for the

constructs of Figure S4B, the 5′ portion of human Zdhhc3 was amplified from purified

human cDNA (courtesy C. Jan) using primers 5′-CGA TTG ACT GAG TCG CCC GGA

TCC GCG TCA TCA ACC TGC GCG G-3′ (forward) and 5′-CTC CTC GCC CTT GCT

CAC ACC AGA ACC ACC ACA GGC GAT GCC ACA GCC-3′ (reverse). Amplified

fragments were purified by polyacrylamide gel electrophoresis, cloned using the Zero

Blunt TOPO kit (ThermoFisher), and sequenced. Specific point mutants (indicated in

Figure S4) were generated by PCR using primers harboring appropriate mismatches.

Cloned fragments were then amplified by PCR and fused with an eGFP sequence and

inserted into the LeGO-iC2 vector (Weber et al., 2008) using HiFi assembly (New

England Biolabs). The CDS of human MMP24-AS1 was amplified from human genomic

DNA using PCR primers 5′-CGA TTG ACT GAG TCG CCC GGA TCC GAG ACC

ATG GGG GCT CAG CTA AG-3′ (forward) and 5′-GCT CCT CGC CCT TGC TCA

CAC CGG AGC CAC CGG TGT ACC AGG AAG TGC AGG CGA TG-3′ (reverse) and

inserted into the LeGO-G2 vector using HiFi assembly. Inserts in all final constructs were

validated by sequencing.

For fluorescence measurements and microscopy, HEK 293T or HeLa cells were

grown in DMEM medium with high glucose, supplemented with glutamine,

penicillin/streptomycin, and 10% FBS. Constructs were transfected using TransIT-LT1

(Mirus) two days prior to data collection. For the Fxr2 and Zdhhc3 experiments (Figure

S4), constructs were transfected into 293T cells, and eGFP and mCherry fluorescence

were quantified using a BD Bioscience LSR-II flow cytometer. Imaging of eGFP-tagged

MMP24-AS1 was performed in HeLa cells co-transfected with ER-mCherry or Golgi-

mCherry constructs as indicated, cultured on a 24-well #1.5H glass bottom plate

(Cellvis). Cells were imaged live on a widefield epifluorescence microscope using a

Lambda LS illuminator (Sutter), 100X Nikon Plan Apo VC 1.4 oil objective, and Andor

Clara CCD camera, controlled with Micro-Manager software (Edelstein et al., 2010).

SUPPLEMENTAL REFERENCES

Cox, J., and Mann, M. (2008). MaxQuant enables high peptide identification rates,

individualized ppb-range mass accuracies and proteome-wide protein quantification. Nat.

Biotechnol. 26, 1367-1372.

Cox, J., Neuhauser, N., Michalski, A., Scheltema, R.A., Olsen, J.V., and Mann, M.

(2011). Andromeda: a peptide search engine integrated into the MaxQuant environment.

J. Proteome Res. 10, 1794-1805.

Csardi, G., and Nepusz, T. (2006). The igraph software package for complex network

research. InterJournal, Complex Systems 1695, 1-9.

Edelstein, A., Amodaj, N., Hoover, K., Vale, R., and Stuurman, N. (2010). Computer

control of microscopes using µManager. Current Protocols in Molecular Biology 14.20.

1-14.20. 17.

Farrell, C.M., O'Leary, N.A., Harte, R.A., Loveland, J.E., Wilming, L.G., Wallin, C.,

Diekhans, M., Barrell, D., Searle, S.M., Aken, B., et al. (2014). Current status and new

features of the Consensus Coding Sequence database. Nucleic Acids Res. 42, D865-72.

Haas, B.J., Delcher, A.L., Mount, S.M., Wortman, J.R., Smith, R.K.,Jr, Hannick, L.I.,

Maiti, R., Ronning, C.M., Rusch, D.B., Town, C.D., Salzberg, S.L., and White, O.

(2003). Improving the Arabidopsis genome annotation using maximal transcript

alignment assemblies. Nucleic Acids Res. 31, 5654-5666.

Hinrichs, A.S., Karolchik, D., Baertsch, R., Barber, G.P., Bejerano, G., Clawson, H.,

Diekhans, M., Furey, T.S., Harte, R.A., Hsu, F., et al. (2006). The UCSC Genome

Browser Database: update 2006. Nucleic Acids Res. 34, D590-8.

Huang, D.W., Sherman, B.T., and Lempicki, R.A. (2008). Systematic and integrative

analysis of large gene lists using DAVID bioinformatics resources. Nat. Protoc. 4, 44-57.

Jovanovic, M., Rooney, M.S., Mertins, P., Przybylski, D., Chevrier, N., Satija, R.,

Rodriguez, E.H., Fields, A.P., Schwartz, S., Raychowdhury, R., et al. (2015). Dynamic

profiling of the protein life cycle in response to pathogens. Science 347, 1259038.

Lin, M.F., Jungreis, I., and Kellis, M. (2011). PhyloCSF: a comparative genomics method

to distinguish protein coding and non-coding regions. Bioinformatics 27, i275-82.

Mertins, P., Qiao, J.W., Patel, J., Udeshi, N.D., Clauser, K.R., Mani, D., Burgess, M.W.,

Gillette, M.A., Jaffe, J.D., and Carr, S.A. (2013). Integrated proteomic analysis of post-

translational modifications by serial enrichment. Nat. Methods 10, 634-637.

Mertins, P., Udeshi, N.D., Clauser, K.R., Mani, D.R., Patel, J., Ong, S.E., Jaffe, J.D., and

Carr, S.A. (2012). iTRAQ labeling is superior to mTRAQ for quantitative global

proteomics and phosphoproteomics. Mol. Cell. Proteomics 11, M111.014423.

Mertins, P., Yang, F., Liu, T., Mani, D.R., Petyuk, V.A., Gillette, M.A., Clauser, K.R.,

Qiao, J.W., Gritsenko, M.A., Moore, R.J., et al. (2014). Ischemia in tumors induces early

and sustained phosphorylation changes in stress kinase pathways but does not affect

global protein levels. Mol. Cell. Proteomics 13, 1690-1704.

Pei, B., Sisu, C., Frankish, A., Howald, C., Habegger, L., Mu, X.J., Harte, R.,

Balasubramanian, S., Tanzer, A., and Diekhans, M. (2012). The GENCODE pseudogene

resource. Genome Biol. 13, R51.

Picard, J. (1976). Maximal closure of a graph and applications to combinatorial problems.

Manage. Sci. 22, 1268-1272.

Rappsilber, J., Mann, M., and Ishihama, Y. (2007). Protocol for micro-purification,

enrichment, pre-fractionation and storage of peptides for proteomics using StageTips.

Nat. Protoc. 2, 1896-1906.

Shalek, A.K., Satija, R., Adiconis, X., Gertner, R.S., Gaublomme, J.T., Raychowdhury,

R., Schwartz, S., Yosef, N., Malboeuf, C., and Lu, D. (2013). Single-cell transcriptomics

reveals bimodality in expression and splicing in immune cells. Nature 498, 236-240.

Sievers, F., Wilm, A., Dineen, D., Gibson, T.J., Karplus, K., Li, W., Lopez, R.,

McWilliam, H., Remmert, M., Soding, J., Thompson, J.D., and Higgins, D.G. (2011).

Fast, scalable generation of high-quality protein multiple sequence alignments using

Clustal Omega. Mol. Syst. Biol. 7, 539.

Spouge, J., Wan, H., and Wilbur, W. (2003). Least squares isotonic regression in two

dimensions. J. Optim. Theory Appl. 117, 585-605.

Stern-Ginossar, N., Weisburd, B., Michalski, A., Le, V.T., Hein, M.Y., Huang, S.X., Ma,

M., Shen, B., Qian, S.B., Hengel, H., et al. (2012). Decoding human cytomegalovirus.

Science 338, 1088-1093.

Waterhouse, A.M., Procter, J.B., Martin, D.M., Clamp, M., and Barton, G.J. (2009).

Jalview Version 2--a multiple sequence alignment editor and analysis workbench.

Bioinformatics 25, 1189-1191.

Weber, K., Bartsch, U., Stocking, C., and Fehse, B. (2008). A multicolor panel of novel

lentiviral “gene ontology” (LeGO) vectors for functional gene analysis. Molecular

Therapy 16, 698-706.